Drums · mean of six sets
0.860
Striker 2.0 10290.860
Striker 1.10.834
Striker2.0
Release Striker 2.0 1029
One neural network that listens to a whole song and finds its drum hits, beats and bass.
Try it in the clubOne network. The whole song.
Striker 2.0 listens to music for LightningStrike, a lighting app. Every 10 ms it says when and how hard the drums hit, where the beats and bars fall, when a bass note starts and which instruments play. It replaces a whole chain of separators and models.
Music in. Rhythm out.
The whole song
Striker 2.0 One network130 M parameters
Every 10 ms
-
Drums
Kick, snare, hi-hat, cymbals and toms, with how hard each hit lands.
-
Beats and bars
Every beat, and the first beat of every bar.
-
Bass
The moment each new bass note starts.
-
Instruments
When drums, bass, vocals and everything else play.
For illustration: the outputs drawn for the loop at the top of the page. They are not real model output.
The frontier
At the frontier.
Striker 2.0 1029 moves the frontier where the comparisons below show it: drum transcription on public benchmarks it never trained on, electronic music, and bass onsets. All of it as one network. The frontier in every category is what Striker 2.1 aims for.
Drums · ENST
0.872
Striker 2.0 10290.872
Weber et al., 20250.80
Drums · RBMA13, electronic
0.767
Striker 2.0 10290.767
Weber et al., 20250.67
Drums · IDMT
0.985
Striker 2.0 10290.985
Noise-to-Notes, 20260.949
Downbeats · Raveform, electronic
0.929
Striker 2.0 10290.929
Beat This!0.742
Bass · MoisesDB
0.917
Striker 2.0 10290.917
Separator and detector0.854
Every number belongs to the release Striker 2.0 1029. It is an F-measure from 0 to 1 on test songs the model never trained on: a 1 means every hit found and none invented. A hit or bass note counts within 50 ms of the real one, a beat within 70 ms. Below Striker is always the best comparable system on the same set.
Why Striker 2.0 is better.
What sets it apart from what came before.
One network, not a chain
One network handles drums with their loudness, beats, downbeats, bass and instruments at once, and it is built for an ordinary computer. The earlier pipeline needed sound separators and several models for that.
Ahead of published results
On ENST, RBMA13 and IDMT drums it leads every comparable published model, without ever training on those sets.
Strong on electronic music
Its biggest leads are on electronic music, which was the earlier chain’s weakest spot.
Measured honestly
It never trains on the test sets, and its results were predicted in writing before training.
Three speeds of listening.
Whether a hit is a snare or a clap, and whether a beat starts a bar, only shows in the bars around it. So the network listens at three rates at once and answers 100 times a second.
100× per second
The exact time of every hit.
25× per second
The bars around each hit.
12.5× per second
The deepest layers, where most of the 130 M parameters live.
What comes next.
Striker 2.0 is made for the LightningStrike app and is not public. Work on it continues.